The Death of SaaS, The Dawn of Agents With Fin

SaaS之死与Fin智能体时代的黎明 Brian Donohue 认为，AI 智能体的出现标志着 SaaS 行业迎来“不适应便消亡”的转折点，并呼吁企业摆脱渐进式的 Copilot 工具，转向完全自主的问题解决系统。本演讲阐明了为何必须用真正的自动化率（介入率与解决率的乘积）以及基于对话记录层面的 AI 质量评分，取代首次响应时间和 CSAT 等传统 CX 指标。要从简单提供信息的 FAQ 进阶到可修改数据库的操作，必须以严谨的软件开发思维来对待提示工程。归根结底，客户支持团队正在从工单管理者结构性转型为软件构建者，成为企业生产级 AI 的一线先锋。 SaaS 行业已迎来向 AI Agent 转型的关键拐点；Intercom 旗下旗舰客服智能体 Fin 目前已拥有超 6,000 家付费客户（包括 Anthropic），验证了自主智能体替代传统软件交互的必然趋势。 (0:04) 企业应警惕仅能带来约 30% 辅助效率提升的“cope-pilot”副驾驶陷阱，转向可端到端直接解决问题的自主 Agent；Fin 客户解决率中位数达 60%–70%，而在自研与采购的权衡中，10% 的解决率差距足以抹平自研软件带来的成本节省。 (3:40) 传统 CX 指标全面失效：首次响应时间（FRT）彻底过时，人工平均处理时间（AHT）因承接复杂问题而自然上升；真实自动化率由介入率乘以解决率决定（业内平均为 47% × 62% = 29%），脱离介入率的高解决率仅是虚荣指标。 (8:49) 客服团队的角色正在发生根本变革，从被动的工单排队处理者转型为配置提示词、优化知识库及部署复杂读写操作的软件构建者，并成为整个企业验证生产级 AI 价值的先锋标杆。 (14:38)

Intercom VP of Product Brian Donohue argues that SaaS companies must pivot from human-assisting copilots to autonomous AI agents capable of resolving support tickets end-to-end. Drawing on deployment data from Intercom's Fin agent, Donohue outlines the operational, metric, and organizational shifts required as support teams transition into software builders. Autonomous Agents Versus Copilot Incrementalism (3:40) Donohue warns SaaS leaders against settling for copilot features, which Intercom co-founder Des Traynor termed the 'cope-pilot' phase. Copilots alter human workflows for an unreliably measured efficiency gain of roughly 30% while leaving human representatives as the core operational bottleneck. In contrast, fully autonomous agents directly resolve customer tickets end-to-end without human intervention. Across more than 6,000 paying customers using Intercom's Fin agent, the median customer achieves a 60% to 70% resolution rate, with approximately 15% exceeding 80%. Donohue references an analysis by Intercom CTO Darragh Curran indicating that raw resolution rate is the primary economic driver in build-versus-buy decisions. A 10% gain in resolution capability comfortably offsets the platform fee of $0.99 per resolution compared to a hypothetically free in-house build with a lower 40% resolution rate, because unhandled tickets sustain higher human labor expenses. Redefining Support Metrics and True Automation Rates (7:13) Traditional customer support metrics lose relevance in an agent-led model. First Response Time becomes obsolete because AI replies immediately, while human Average Handling Time actually rises as simple inquiries are automated and only complex escalations reach staff. Donohue also notes that traditional CSAT surveys suffer from low response rates of around 10%, introducing heavy selection bias. Instead, organizations should use AI to evaluate 100% of conversation transcripts, which consistently scores automated responses higher due to adherence to policy without emotional fatigue. Furthermore, high resolution rates can mask low business impact if query routing is overly restrictive. Donohue highlights the relationship between resolution rate and involvement rate—the share of total volume exposed to the agent. Across Fin's customer base, an average 62% resolution rate combined with a 47% involvement rate yields an actual automation rate of 29%. Maximizing impact requires organizations to elevate involvement rates, with leading adopters pushing routing to 90%–98% of inbound volume. Moving Beyond FAQs to Actions and API Integrations (10:54) Support queries fall into three distinct tiers: informational read-only queries, personal read-only lookups requiring backend database access, and read-write actions that directly modify database records. While read-write operations deliver the highest business value, they also present the greatest operational risk. Executing actions requires configuring natural language triggers and detailed instructions. Donohue notes that while natural language has effectively become the programming language, teams still require an engineering mindset to write predictable workflows. Intercom's internal research identified four key blockers to adoption: unwritten institutional processes, steep platform learning curves, inconsistent prompt engineering, and absent or poorly structured internal APIs. Consequently, advanced action capabilities remain under managed availability. Support Teams as Frontline Software Builders (14:38) Customer experience departments are moving away from manual queue triage and human capacity management toward an ongoing software development lifecycle. Support teams now operate a continuous loop encompassing conversation analysis, content and guidance training, testing, and multi-channel deployment. In this setup, support staff operate as builders who design the frontline customer journey and maintain the software defining it. Practitioners achieve performance gains by maintaining guidance prompts—Intercom's internal team manages over 80 active snippets—and restructuring internal documentation so that the AI agent serves as the primary reader. Teams leverage automated suggestions with roughly a 70% acceptance rate for content updates. Because support operations generate clear, measurable resolution metrics, organizations like Consensys have used their CX teams as the internal pioneer and blueprint for broader enterprise AI implementation. Donohue positions customer support as the most quantifiable proving ground for enterprise AI. However, successfully scaling autonomous agents requires organizations to move beyond basic FAQs by resolving backend API gaps, formalizing unwritten institutional workflows, and actively managing natural language instructions as production code.

Intercom 产品副总裁 Brian Donohue 提出，SaaS 行业正迎来被 AI Agent 重塑的转折点；企业不应满足于仅带来有限效率增量的人工辅助工具（Copilot），而必须将端到端自主解决率作为核心经济杠杆，并将客服团队重塑为构建与维护智能体软件的核心力量。 摆脱 Copilot 局限与解决率经济学 (3:40) Brian Donohue 指出，许多 SaaS 团队寄希望于为人工坐席配备 Copilot 界面，但此类工具通常仅能带来约 30% 且难以量化的人工作业效率提升，Intercom 联合创始人 Des Traynor 将其戏称为“应对阶段”（cope-pilot）。相比之下，能够直接处理客户诉求且无需人工介入的自主 AI Agent 才是带来结构性变革的核心。根据拥有超 6,000 家付费客户的 Fin 运营数据，客户解决率中位数介于 60% 至 70% 之间，约 15% 的客户解决率超过 80%，Anthropic 等 AI 机构均采用了该方案。 在自研与采购（Build vs. Buy）的权衡上，Intercom 首席技术官 Darragh Curran 的模型表明，解决率是决定商业价值的核心杠杆。假设企业自建 Agent 系统的边际软件成本为零，但解决率仅有 40%；而 Intercom Fin 即使按每次解决收取 0.99 美元，若解决率达到 50% 的保守基准，其高出 10% 的解决率所节省的人力工时成本，足以抵消软件采购支出，使高解决率方案的综合净收益显著优于看似免费但解决率偏低的自建系统。 颠覆客服传统指标与计算真实自动化率 (7:13) 在 Agent 优先的模式下，传统指标正面临失效。首次响应时间（FRT）因 AI 的即时应答而失去区分度；人工坐席的平均处理时长（AHT）则由于常规简单问题被分流、仅留存复杂疑难与升级个案而必然上升。此外，传统客户满意度（CSAT）调查因仅约 10% 的回复率而存在严重抽样偏差，Intercom 主张由 AI 对 100% 对话记录进行客观全量质检。评估显示，AI 凭借即时、稳定且严格遵循业务规范的输出，在全量评审中得分普遍高于传统人工服务。 衡量成效的关键在于“真实自动化率”，即“介入率（Involvement Rate）× 解决率（Resolution Rate）”。如果团队单方面追求 60% 的高解决率，却仅向 Agent 开放 10% 的进线流量，其实际自动化率仅有 6%。目前 Fin 客户的平均解决率为 62%，平均介入率为 47%，折合真实自动化率约为 29%；而领先企业（如 Money Group）已将介入率推高至 90%–98%，从而最大化释放自动化带来的运营规模效应。 跨越问答层级与自然语言工程化陷阱 (10:54) 客户服务交互可划分为三个层级：一是仅读取通用知识库的“信息类”查询；二是调取后端数据库的“个人数据类”只读查询（如查询余额或物流进度）；三是直接修改数据库的“操作类”（Action，如修改地址或退换货），后者价值最高但也伴随最高风险。 Brian Donohue 强调，“自然语言已成为新的编程语言，但这本质上依然属于编程”。阻碍企业落地高级操作任务的核心瓶颈在于：未成文的内部业务流程、平台学习成本、提示词工程不一致，以及内部 API 缺失或结构混乱。为此，高阶操作任务目前多采取受管模式逐步推行；而相对轻量的单行“指导”（Guidance）规则已能产生显著收益，例如有电商客户仅凭借几句提示指令，便将 Fin 转变为高效的线索捕获与销售转化工具。 客服团队从工单管理重塑为软件构建者 (14:38) 随着 AI Agent 承接一线交互，客服团队的工作重心从传统的工单排班与人工分流，转变为负责智能体软件全生命周期的构建者，围绕“洞察（Insights）—训练（Train）—测试（Test）—部署（Deploy）”四步闭环持续运作。 在此体系下，支持团队逐步转向对话设计师与知识架构师角色：通过分析人工介入的漏网问题补充知识库，重构文档以满足 AI 机器阅读标准，并持续调优自然语言提示词。在 Intercom 内部实践中，团队维护着 80 余条活跃微调指令，采纳了约 70% 的自动化内容建议，并上线 28 项单行微调与 8 个复杂任务；一线从业人员对 Agent 行为的主动调优，已成为推动解决率持续提升的关键动能。 客服部门作为企业 AI 落地的度量标杆 (17:51) 相较于内部研发部门常面临借助拉取请求（PR）数量等粗糙指标来证明“人效翻倍”的考核压力，客服场景具备直接面向客户的解决率和清晰的 ROI，因而在企业推进 AI 落地时展现出最高的确定性与可见度。 领先的客服组织正逐步成为全公司的“AI 先锋”。以 Consensys 等企业为例，客服团队率先将大语言模型与 Agent 交互系统结合并成功投入生产运行，促使其他业务部门主动前来咨询并借鉴其构建、监控与扩展 AI 系统的工程经验。这不仅提升了支持效率，更在更广泛的层面上重塑了人机交互界面的未来形态。 Brian Donohue 总结认为，AI Agent 正在加速终结传统 SaaS 的交互与交付逻辑，客服领域则是首个验证其业务回报的试金石。然而，企业要实现从被动问答到全流程自动化的跃升，仍需攻克底层 API 规范化、隐性业务规则显性化等工程挑战；企业只有将支持人员重塑为长效维护智能体的软件构建者，并追求高介入率与高解决率的双轮驱动，方能在 Agent 时代建立持久壁垒。

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布，主讲人为 Intercom 旗下 AI 客服产品 Fin 的产品副总裁 Brian Donohue。视频简介指出，AI Agent 时代已经全面到来，本场演讲旨在抛开空洞的行业炒作，为企业提供一份将现有产品转型为 Agent 的实用落地实操指南。 ## TL;DR Brian Donohue 指出传统 SaaS 正面临向 AI Agent 转型的生死关头，告诫企业切勿停留在 Copilot 这类只能带来微小效率改善的“面包屑”陷阱中。企业必须彻底重构原有的客服评估体系，用 AI CX Score、介入率与端到端自动化率替代传统 CSAT 与工单时长。随着 AI 从处理简单信息查询向执行高风险动作（如修改数据库）演进，客服团队的角色也正在从管理人力的坐席转向构建与微调 Agent 的“软件构建者”。由于成效极易量化，客户体验（CX）正成为企业全盘落地 AI、重塑人机交互方式的最佳先锋阵地。 ## Atmosphere & style 演讲风格务实、犀利且充满紧迫感，开篇引用科技大佬与灭霸的名言直击传统 SaaS 的存亡危机，迅速抓住听众注意力。随后，主讲人抛弃宏大叙事，通过 Intercom 详实的后台数据与真实踩坑教训，以一种“老兵传授实操经验”的坦诚口吻层层拆解落地难点。整体节奏紧凑、逻辑严密，既有直面业务转型的硬核洞察，又为身处 AI 焦虑中的管理者指明了清晰且充满希望的转型路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) AI Agent 时代的到来正将传统 SaaS 企业推向“要么适应、要么灭亡”的关键抉择点，全面拥抱 Agent 成为重构竞争力的核心路径。 - [02:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=170s) 警惕 Copilot（副驾驶）模式带来的渐进式效率陷阱，企业切勿自满于难以量化的边际提升，而应追求端到端全自动闭环的 AI Agent 模式。 - [02:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=170s) 即使假设自研系统成本为零，解决率更高的商业化 AI Agent 所带来的实际业务价值也远超自研低解决率系统所节省的开销。 - [07:15](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=435s) 传统客服指标全面失效：首次响应时间失去意义，人工处理时长因接手疑难杂症而自然上升，AI 全量会话质检的 CX Score 取代了仅有 10% 回复率的 CSAT 调查。 - [08:49](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=529s) 真实的端到端业务自动化率等于“介入率”（组织开放度与集成范围）乘以“解决率”（AI 模型能力），单一关注解决率容易产生认知偏差。 - [10:54](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=654s) AI 进线处理正从简单的通用信息问答、数据库读取查询，深入到高业务风险与高价值的“数据库写入与动作执行”深水区。 - [10:54](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=654s) 虽然英语成为了配置 AI Agent 的新编程语言，但本质依然是软件工程，缺乏清晰文档化 SOP 和提示词工程思维会导致复杂任务落地受阻。 - [14:35](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=875s) 客服团队正在从被动的“人员管理与工单处理者”转型为“软件构建者”，形成涵盖洞察、训练、测试、部署的高频软件迭代闭环。 - [14:35](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=875s) 企业需要将 AI Agent 视为内部知识消费的“第一受众”，并随之设立知识内容架构师、对话设计师等新型专职 AI 岗位。 - [17:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=1070s) 相比研发等难以精准量化效率的部门，客户体验（CX）拥有最清晰且直接的成效衡量体系，是引领整个企业 AI 落地的最佳先锋阵地。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) AI Agent 时代的到来与 SaaS 的生死存亡 Brian Donohue 开场直奔主题，引用了科技界领袖对 AI 的宏大预测。DeepMind 首席执行官 Demis Hassabis 曾表示，回顾历史，AI Agent 将成为人类历史上最具变革性的发明；英伟达（NVIDIA）CEO 黄仁勋（Jensen Huang）也曾预言，AI Agent 的普及曲线将如同移动互联网革命一般——“起初悄无声息，转眼间无处不在”（"Invisible at first and everywhere"）。Brian 坦言，整个行业正充斥着这股浪潮，而打造了 AI 客服 Agent 产品 **Fin** 的 Intercom 团队坚信，对于现有的 SaaS 企业而言，这正是一个“要么适应，要么灭亡”（"adapt or die moment"）的关键时刻。 ```mermaid flowchart LR accTitle: SaaS与AI Agent转型路径 subgraph S["当前抉择"] A["传统 SaaS 模式"] --> B{"是否拥抱 Agent?"} end subgraph O["转型结果"] B -- "全力投入" --> C["品类领军者<br/>营收暴增 / 登上福布斯"] B -- "迟疑逃避" --> D["逐渐边缘化<br/>最终被市场淘汰"] end ``` 在过去两年半的时间里，Intercom 基于这种信念对其整个公司进行了彻底的重塑。Brian 借用《复仇者联盟》中灭霸（Thanos）的名言幽默地说道：“无论你多么害怕，无论你如何逃避，Agent 时代终将到来”（"dread it, run from it, agents will arrive all the same"）。面对这场变革，乐观的路线是抓住机遇成为品类领导者，收获数以千万计的收入并登上《福布斯》杂志封面；而对于需要危机感驱动的人来说，如果失败就只能眼睁睁滑向无足轻重的边缘。Brian 表示，本次分享将抛开空洞的行业炒作，聚焦于客户体验（CX）领域，为 CX 负责人和企业管理者提供五个实用的战略认知： 1. **真正的业务转型已经触手可及，切勿满足于捡拾面包屑**（Don't settle for the breadcrumbs）。 2. **必须彻底重构传统的评估指标体系**（Rethink your metrics）。 3. **AI Agent 绝不仅用于处理简单问题，但要让其胜任复杂任务绝非易事**。 4. **CX 角色必须转变：客服团队正在转型为软件构建者**（Software builders）。 5. **CX 团队应当成为全公司 AI 转型的灯塔与先锋**。 --- ### [02:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=170s) Fin 的行业验证与 Copilot 的“面包屑”陷阱 在深入探讨之前，Brian 简要展示了 Intercom 在 AI Agent 领域的实力背书：Fin 作为专为客户支持构建的 AI Agent，已被 G2 连续三次评为顶级 AI Agent，拥有超过 6,000 家付费客户。Bessemer Venture Partners 甚至评价其为“当下被最严重低估的 AI 公司”；最令团队自豪的客户则是大模型先锋 Anthropic，他们也在一线支持业务中深度使用 Fin。 ```mermaid flowchart TD accTitle: Copilot与AI Agent的对比 subgraph Copilot["Copilot 模式（面包屑）"] C1["辅助人工客服草拟回复"] C2["需要人类改变工作流程"] C3["效率提升有限且极难量化 (~30%)"] end subgraph Agent["AI Agent 模式（真正转型）"] A1["端到端全自动闭环解决"] A2["直接解放人力接管复杂问题"] A3["整体解决率跨越式提升 (50%-80%+)"] end ``` 谈及核心观点“不要满足于面包屑”，Brian 举例展示了 RB2B 这家公司如何将整个业务深度交由 AI 自动化运转，并在 LinkedIn 上引发热议。随后，他将矛头对准了当前泛滥的“Copilot（副驾驶）”模式： - **Copilot 的局限性**：Intercom 曾在 2024 年 4 月上线了精心打磨、带有炫酷渐变色彩 UI 的 Copilot 产品。虽然它确实是一个有价值的辅助工具，但 Brian 直言有点后悔做它，因为很多企业把 Copilot 当作了偷懒的“拐杖”。Copilot 本质上是在做渐进式的效率优化，试图改善人类的工作方式，但声称的 30% 效率提升极其难以衡量，根本没有触及问题的核心。 - **“Cope-pilot”的心理防御**：Intercom 联合创始人 Des Traynor 将其嘲讽为“Cope-pilot（心理应对驾驶员）阶段”——企业用 Copilot 辅助团队，然后自欺欺人地告诉自己“一切都不会发生根本改变”。 - **Agent 带来的真实飞跃**：Intercom 后台 6,000 家客户的数据显示，AI 驱动的平均问题解决率（Resolution Rate）在两年半内通过数百次 A/B 测试实现了持续爬升。在客群分布中，中位数落在 60% 至 70%，更有约 15% 的客户实现了超过 80% 的端到端全自动解决率。 Intercom 首席技术官（CTO）Darragh 曾算过一笔经济账：即使假设自研 AI Agent 的研发成本为零，如果自研系统只能达到 40% 的解决率，而使用收费 0.99 美元/次解决的 Fin 能达到 50% 的解决率，那么这多出来的 10% 解决率所产生的实际业务价值，也远远超越了所谓“免费”产品的节省额度。 --- ### [07:15](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=435s) 重塑 CX 指标体系：告别 CSAT，迎来 AI 驱动的 CX 评分 随着 AI Agent 成为一线支持的主力，传统的客服考核指标体系彻底失效，企业必须重新定义衡量的维度： | 传统指标 | 为什么在 Agent 时代失效 / 逆向变化 | 替代/核心指标 | |:--- |:--- |:--- | | **首次响应时间 (FRT)** | AI Agent 可以做到毫秒级即时响应，该指标直接失去衡量意义。 | **解决率 (Resolution Rate)**：衡量问题是否得到端到端彻底闭环。 | | **平均处理时长 (AHT)** | AI 接管了海量常规问题，人类客服处理的只剩棘手、愤怒的高难度长尾案例，导致**人类 AHT 自然上升**。 | **自动化分流率 (Volume Rate)**：衡量 AI 分担了多少整体工单压力。 | | **客服代表生产力** | 传统产出模式极难在人机混合协作场景下精准量化。 | **CX Score (AI 体验评分)**：由 AI 全量审阅对话并多维度打分。 | | **传统满意度调查 (CSAT)** | 用户问卷回复率通常仅有 10% 左右，样本严重偏差且失真。 | **全量会话质检分析**：AI 逐行研读 100% 的对话实录并评估服务质量。 | Brian 重点强调了由 AI 驱动的 **CX Score**。在过去，企业依赖回复率仅 10% 的问卷调查来评估 CSAT，这种方式既低效又容易失真。在 Agent 时代，AI 不仅负责提供服务，还在负责评估服务——它能够阅读 100% 的完整对话实录，客观评估每一次沟通的交互质量，既监督 AI Agent 自身，也同时评估人类员工的服务表现。有趣的是，在 AI 的全量质量评估中，人类客服的表现往往略逊于预期，而 AI Agent 反而评分更高，这进一步证明了全面转向 Agent 模式的必要性。 ### [08:49](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=529s) 评估指标的真相：介入率、解决率与真实的自动化率 在衡量 AI Agent 的实际业务价值时，许多客户往往陷入了只关注单一指标的误区。团队此前在客户绩效报告中提供了直观的可视化漏斗图，展示跨渠道的请求流向以及 AI 与人工的分配情况。然而，许多客户体验（CX）团队过度锚定于“解决率”（Resolution Rate），而忽视了另一个至关重要的核心指标——“介入率”（Involvement Rate）。所谓介入率，是指在所有进入系统的客户请求总量中，系统究竟给了 AI Agent 多少次尝试处理的机会；而解决率则是在 AI 获得尝试机会的前提下，成功闭环解决且无需转交人工团队的比例。 如果一个团队对外宣称其 AI 拥有 60% 的解决率，听起来表现不错，但若该团队的介入率仅有 10%，两者相乘后的实际端到端自动化率仅有可怜的 6%。根据 Fin 付费客户群体的真实统计数据，其平均解决率为 62%，平均介入率约为 47%，最终折算出的全量自动化率（Automation Rate）约为 29%。虽然承接并解决全渠道近三分之一的业务量已经是一个很好的起点，但 29% 与 62% 之间存在着巨大的认知落差。从本质上讲，解决率衡量的是 Fin 作为 AI 工具本身的模型表现，而最终的自动化率衡量的则是客户整个组织的业务落地与流程改造能力。在实践中，像 Money Group 这样的标杆客户已经能够将介入率推高至 98%，从而释放出极高的业务自动化价值。 ```mermaid flowchart LR accTitle: CX Automation Metrics Formula accDescription: Flow showing how Total Inbound Volume is filtered by Involvement Rate and Resolution Rate to calculate the true Automation Rate. subgraph Funnel ["全量客户请求漏斗"] A["全部进线请求<br/>(Inbound Volume)"] -->|"介入率 (47%)<br/>组织赋权与覆盖面"| B["Fin 介入处理<br/>(Handled by AI)"] B -->|"解决率 (62%)<br/>模型与知识能力"| C["成功闭环解决<br/>(Resolved)"] end subgraph Outcome ["最终业务指标"] C --> D["真实自动化率 (29%)<br/>(Automation Rate)"] end ``` - **解决率（Resolution Rate）**：在 AI 介入的对话中成功解决、未流转到人工的比例（平均值约 62%），体现 AI 自身能力。 - **介入率（Involvement Rate）**：全部进线请求中分流给 AI 尝试处理的比例（平均值约 47%），体现组织的开放程度与系统集成范围。 - **实际自动化率（Automation Rate）**：计算公式为 $\text{介入率} \times \text{解决率}$（平均值约 29%），是真正反映业务降本增效的核心数字。 --- ### [10:54](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=654s) 攻克深水区：从简单问答到高风险的“数据库写操作” AI Agent 的应用绝不仅限于处理低难度的边缘问题。尽管业界普遍认为“AI 负责简单事务，人类处理复杂难题”，但这仅仅是第一步。当前的 AI 已经完全有能力执行复杂任务，只是这需要投入更多的配置与打磨时间。通常可以将进线对话划分为三个梯度： 1. **信息类查询（Informational Queries）**：无论谁来询问答案都完全一致，通常直接索引帮助中心文档即可作答，处理阻力最小； 2. **个性化查询（Personal Queries）**：需要连接后台读取数据库（如“查询我的账户余额”或“查询物流状态”）； 3. **动作执行（Actions）**：直接对数据库进行写入或修改（如“更改收货地址”、“取消订单”或“申请退款”），这也是业务风险最高、价值最大的深水区。 ```mermaid flowchart TD accTitle: Three Buckets of Support Queries accDescription: Classification of support volume into Informational, Personal, and Action queries showing increasing risk and complexity. subgraph Queries ["客户服务请求的三大分类"] Q1["信息类查询 (Informational)<br/>读取通用文档 / 统一回复"] Q2["个性化查询 (Personal)<br/>读取数据库 (如：查询余额/状态)"] Q3["动作执行类 (Actions)<br/>写入数据库 (如：修改订单/退款)"] end Q1 -->|难度与风险递增| Q2 Q2 -->|难度与风险递增| Q3 ``` 在 Fin 后台配置这些复杂任务时，管理员可以通过自然语言触发器（Natural Language Trigger）定义场景（例如“针对破损订单的处理流程”），随后使用自然语言编写具体执行指令。然而，这种表面的低门槛具有很强的欺骗性——"English is now the programming language, but you're still doing programming."（英语虽然变成了编程语言，但你本质上依然在做编程）。许多企业团队因为缺乏工程思维，在自然语言提示词配置中频繁踩坑，出现“进一步退两步”的困境。 导致企业在复杂任务落地受阻的核心原因包括： - **内部流程未文档化**：大量关键业务流程只存在于资深员工的大脑中，缺乏清晰的 SOP。 - **系统学习与适应曲线**：即使配置界面追求低门槛，操作人员仍需理解 AI Agent 的行为逻辑。 - **提示词工程与行为确定性**：CX 团队必须掌握让大模型表现出可预测行为的 Prompt 调试能力。 - **内部 API 接口限制**：后端接口不规范或权限不开放是推进动作类 Agent 的主要技术卡点，这也促使 Fin 暂时将此类深度功能设为受控开放（Managed Availability）模式以协助客户推进。 与此同时，简单的指令微调也能带来意想不到的惊喜。例如某电商客户仅在 Fin 中补充了一条单行指导指令（Guidance），就成功让 Agent 在独立站上开始自主承接销售线索（Leads）转化。 --- ### [14:35](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=875s) CX 团队的角色重塑：客服专家转型为软件构建者 随着 AI Agent 成为团队中吞吐量最大的“超级队友”，客服团队的核心职责正发生结构性剧变：支持团队不再仅仅管理人力坐席，而是转变为了“软件构建者”（Software Builders）。整个运作机制变成了一个高频迭代的软件开发闭环：从数据洞察（Insights）发现盲区，到知识与规则训练（Train，包括文档内容、Guidance 提示词与 Task 任务），再到全面测试验证（Test），最后部署至各渠道（Deploy），周而复始。 ```mermaid flowchart LR accTitle: Agent Continuous Improvement Loop accDescription: Continuous iteration cycle of building AI Agents: Insights, Train, Test, and Deploy. subgraph Loop ["Agent 软件化构建闭环"] I["洞察 (Insights)<br/>分析进线会话与未解问题"] --> TR["训练 (Train)<br/>优化文档、Guidance与Task"] TR --> TE["测试 (Test)<br/>回归验证与行为校准"] TE --> D["部署 (Deploy)<br/>分发至全渠道上线"] D --> I end ``` 在这一变革下，支持团队内部涌现出了专职的 AI 专家角色： - **对话洞察分析师**：利用 AI 深度分析进线话题分布。系统会自动标记出“Fin 无法回答但人工客服成功解答”的信息类问题，甚至能直接生成建议补充的知识库草稿，人工只需一键点击批准（Approve）。 - **知识与内容架构师（Knowledge Manager）**：专门面向 AI 优化文档结构与表达。在 Intercom 内部，团队维护着大约 80 条针对性的 Guidance 规则，这是历经两年持续迭代精调的成果，属于全职且长期的核心工作。 - **对话设计师（Conversation Designer）**：构建确定性的工作流与多轮交互逻辑。 在 Intercom 自身的客户支持实践中，团队正是通过增设专职的知识库管理人员与对话设计师，将 AI Agent 视为内部知识消费的“第一受众”（First Audience），从而在保证高客户满意度评分（CSAT）的同时，实现了解决率与介入率的大幅飞跃。 ### [17:08](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=1028s) 现代 CX 团队的转型：从管理工单到构建 AI Agent 软件 客户支持团队在优化 AI Agent 时的实际工作成果展现了惊人的执行力与数据表现。在内容建议（Content suggestions）方面，团队采纳率达到了约 70%；同时完成了 28 项单行提示词微调（one-line fine-tuning）以及 8 个全新的 Fin Tasks，用以处理更复杂的业务任务。与此同时，研发（R&D）团队则负责对话式 Fin 产品的底层迭代与大量 A/B 测试，并将这些优化成果直接赋能给一线。这就是面向未来的现代客户体验（CX）团队每天都在进行的真实工作。 过去与未来的分水岭十分明确：过去的 CX 团队主要工作是“管理人员和处理工单”，而未来的团队核心则是“构建与优化 AI Agent”。如今的客户支持团队正在演变成事实上的产品团队（Product teams），他们所调试、配置和优化的 Agent 软件，正在直接定义并交付客户的最终体验。 ```mermaid flowchart LR accTitle: CX Team Evolution subgraph Past["传统模式 (Past)"] A["人工客服"] --> B["工单处理 (Tickets)"] end subgraph Future["未来模式 (Future)"] C["CX 团队转型为产品团队"] --> D["构建与微调 Fin Agent"] D --> E["交付全新客户体验"] end ``` - **内容建议采纳率**：高达约 70% 的优化建议被团队直接采纳实施。 - **配置与调优细节**：团队完成了 28 项提示词微调，并搭建了 8 个全新的复杂 Fin Tasks。 - **角色重塑**：从被动的工单处理者转变为主动构建客户体验软件的“产品开发者”。 --- ### [17:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=1070s) 客户体验（CX）是企业落地 AI 的最佳先锋阵地 许多企业领导者在面对琳琅满目的 AI 场景时往往感到困惑，不知道究竟应该从何处切入。答案其实非常明确：CX 是眼前最成熟、收益最显著的突破口，理应成为整个企业 AI 落地的先锋阵地。虽然研发团队（R&D）内部也面临着巨大考核压力——比如要求工程师实现“2倍生产力提升”，并且只能粗糙地用提交 PR（Pull Requests）的数量来衡量——但大家心知肚明，这并不是一种衡量工程效率的好方法。相比之下，CX 领域拥有行业内最清晰、最容易量化 AI Agent 成效的业务指标。 业界对于这场变革的前景达成了深刻共识：“我们不只是在落地 AI 工具，而是在彻底重新构想人类与软件的交互方式。最先弄懂这一点的企业，获得的不仅是更出色的客服支持，而是截然不同的全新用户体验。” 以 Consensys 为例，其 CX 团队正引领着全公司的 AI 落地应用，其他部门甚至需要向他们取经，学习如何在生产环境中部署 AI。CX 团队完全有能力成为引领整个企业拥抱 AI 的先行者。 - **量化衡量的优势**：相比研发部门用 PR 数量衡量生产力，CX 业务拥有最精准且立竿见影的 AI Agent 评估体系。 - **跨部门示范效应**：诸如 Consensys 的 CX 团队已成为全公司的 AI 标杆，向其他部门输出生产环境的落地经验。 - **全新资源发布**：团队当天上线了帮助企业梳理并规划 Agent 实施路径的实用指南，提供大量落地细节。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) Invisible at first and everywhere — Jensen Huang > [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) adapt or die moment — Brian Donohue > [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) dread it, run from it, agents will arrive all the same — Brian Donohue > [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) Don't settle for the breadcrumbs — Brian Donohue > [00:04](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=4s) Rethink your metrics — Brian Donohue > [02:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=170s) Cope-pilot — Des Traynor > [10:54](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=654s) English is now the programming language, but you're still doing programming. — Brian Donohue > [17:50](https://www.youtube.com/watch?v=DJoa9G2-ZyY&t=1070s) 我们不只是在落地 AI 工具，而是在彻底重新构想人类与软件的交互方式。最先弄懂这一点的企业，获得的不仅是更出色的客服支持，而是截然不同的全新用户体验。 — Brian Donohue ## Mentioned Brian Donohue, Demis Hassabis, DeepMind, Jensen Huang, NVIDIA, Fin, Intercom, G2, Bessemer Venture Partners, Anthropic, RB2B, Des Traynor, Darragh, Money Group, Consensys ## Video description Brian Donohue, VP Product, Fin Abstracts The AI agent era is happening right now. It's not if or when. If you fail to take advantage, you can expect to slide into irrelevance. In this talk I'll share a practical way to think through what it means to turn your offering into an Agent and the steps you’ll need to take to get there. This isn't about parroting hype-speak - it's giving you a playbook for right now [Source](https://www.youtube.com/watch?v=DJoa9G2-ZyY)

[0:04] Jensen Huang: Invisible at first and everywhere [0:04] Brian Donohue: adapt or die moment [0:04] Brian Donohue: dread it, run from it, agents will arrive all the same [0:04] Brian Donohue: Don't settle for the breadcrumbs [0:04] Brian Donohue: Rethink your metrics [2:50] Des Traynor: Cope-pilot [10:54] Brian Donohue: English is now the programming language, but you're still doing programming. [17:50] Brian Donohue: 我们不只是在落地 AI 工具，而是在彻底重新构想人类与软件的交互方式。最先弄懂这一点的企业，获得的不仅是更出色的客服支持，而是截然不同的全新用户体验。

[0:04] Jensen Huang: invisible at first and everywhere [0:04] Brian Donohue: dread it, run from it, agents will arrive all the same [0:04] Brian Donohue: adapt or die moment [3:40] Des Traynor: cope-pilot [10:54] Brian Donohue: English is now the programming language, but you're still doing programming. [17:51] Brian Donohue: We're not just implementing AI tools, we're reimagining how humans interact with software entirely. The companies that figure out this first won't just have better support, they'll have completely different user experiences.

[0:04] Hey folks, uh I'm Brian. Um we'll start [0:08] with the hype, the death of SAS, the [0:11] dawn of agents, and I'm going to [0:12] continue with the hype here. Uh CEO of [0:15] Deep Mind, Alph Go, going back in the [0:17] day, 2014, bought by Google. Uh we'll [0:20] look back and see agents, the most [0:21] transformative transformative invention [0:25] humanity ever created. Choose one of [0:27] your tech leaders uh who's who's gone [0:30] for a breathy one. Let's go for uh [0:32] Jensen. The adoption curve of AI agents [0:34] will mirror the mobile revolution. [0:35] Invisible at first and everywhere. Uh we [0:38] all know the hype. It's all over the [0:40] place here. Um and uh at Intercom, the [0:43] company who built Finn, uh we believe [0:46] this. Um we believe this is uh for for [0:49] SAS businesses an adapt or die moment. [0:52] The agent era is happening now. It's not [0:54] a question of if it is. It's not a [0:56] question of when it is now. We've [0:58] completely reinvented our company over [1:00] the last two and a half years based on [1:02] that conviction. Um, you know, in words [1:05] of Thanos, dread it, run from it, agents [1:08] will arrive all the same. We do believe [1:09] this is inevitable now. And you can kind [1:12] of look at it as like the optimistic [1:13] route. If you're motivated more by that, [1:15] you know, if you nail it, you'd be a [1:17] category leader. You know what? Untold [1:18] millions of revenue, a front page of [1:21] Forbes raise all your rounds. or you [1:23] know if you if you need to be driven by [1:25] fear then it's like if you fail then you [1:27] just slide into irrelevance and won't [1:28] matter anymore. So when we say the the [1:31] death of SAS and the birth of agents, [1:34] that's really what we're talking about [1:35] here. But okay, enough hype. We didn't [1:38] people came to this conference already [1:40] believing at least some of the hype [1:42] here. And what I want to share in this [1:43] brief like 18 minutes left time slot [1:46] here something more practical. Um and I [1:48] want to share. So I'm Brian. I'm a VP of [1:50] product uh at intercom. Spent the last [1:52] two and a half years. I've been there 11 [1:53] years. Last two and a half years focused [1:55] on building Finn, our AI agent. I want [1:57] to share with you and I want I I'll [1:59] focus this talk more on CX and if you're [2:02] not here around CX you can abstract this [2:04] out because I want this to be tangible [2:05] because I think that's what's most [2:06] useful for this for this conference or [2:08] for folks coming here. So here's what [2:10] I'm going to tell you. Uh here's the [2:11] things I wish CX leaders understood. [2:14] One, real transformation is on the [2:16] table. Don't settle for the breadcrumbs. [2:19] Two, you need to rethink your metrics. [2:22] Three, AI agents are not just for the [2:24] easy stuff, but it's not easy to get [2:27] them to do the hard stuff. [2:29] And um it's hard for me to read at this [2:32] sort of angle here. This is a challenge [2:33] for me, but CX roles must change. CX [2:36] teams are now software builders. There's [2:38] a whole transformation happening. And [2:40] overall, I think CX teams are the ones [2:42] who should be the beacon of AI [2:44] transformation in your company. So those [2:46] are five points. Let's see if I can make [2:48] them with the clock ticket. Okay. So why [2:50] why should you believe me? I believe uh [2:52] do I have any credibility up here at [2:53] all? Well, so Finn is the um AI agent [2:57] for customer support that we've built. [2:59] Uh we think it's the uh leading agent [3:01] out there based on here's like obviously [3:02] the sales site, but uh we've got G2 [3:05] giving us credibility here. We voted top [3:07] AI agent three times now. Uh we've got [3:09] over 6,000 paying customers. You can see [3:11] if you're able to see the stuff at the [3:12] bottom, a bunch of AI forward companies [3:14] who have chosen us. Uh this is some [3:16] nicest from Bessemer VP is like hey [3:18] they're the most underhyped AI company [3:20] going there or uh the customer we're [3:23] probably most proud of having Zanthropic [3:25] uh because the credibility that goes [3:27] they use Finn and here is a nice quote [3:29] from Emily kind of given credibility uh [3:33] to what we're doing here and so their [3:34] vote of of Finn is a big vote uh for us [3:37] and the uh the quality of what we're [3:39] building. Okay, so that's credibility [3:40] stuff over. So here's what I want to [3:42] share with you, right? one real [3:43] transformation on the table. Don't eat [3:45] the breadcrumbs. Okay, this is an [3:46] example that I have to go down to [3:48] actually read, but this is like uh you [3:50] guys should check this out. This company [3:52] RB2B, I don't know what they do. They're [3:54] Finn customers. They've like turned off [3:56] their whole company and tried to run it [3:57] all entirely on AI. Go check them out on [3:59] LinkedIn. It's [ __ ] really cool what [4:01] they're trying to do. They are pushing [4:03] the boundaries of what you can do here. [4:04] Now, specifically, this guy is talking [4:06] here. Hey, our AI ticket resolve this. [4:08] And he's like, this is what 2025 looks [4:10] like. You can pretend it's not happening [4:12] or you can adapt. I think they're [4:14] articulating and living this breathy [4:16] talk. They are living it. And here they [4:18] kind of share for fan. Here's the [4:19] numbers they get. I'm not going to go [4:20] through here. They got good numbers. [4:21] Right. Okay. Now, when I'm talking about [4:23] breadcrumbs, this here is a inbox for [4:25] teammates for support. On the right is [4:27] co-pilot. Copilot. There's a thousand of [4:29] these co-pilot products out there. We [4:31] put our heart, sweat, and tears and [4:33] build these. We launched this April last [4:34] year. You can see, you know, it's an AI [4:36] product. It's got to have the gradient [4:37] [ __ ] going on. All that sort of stuff [4:39] there. We fancified the hell out of it. [4:40] But we actually tried to make it a [4:41] really good product as well. And part of [4:43] me almost wishes we didn't. Uh because [4:46] people are like it's the crutch. Like ah [4:48] I'm just going to go for the co-pilot [4:50] and go there. And the the co-pilot is [4:52] useful. It's a valuable product. You've [4:54] got to get your humans to change how [4:56] they work. You got to change how humans [4:57] are working here. And it's like about an [4:59] efficiency game. We measure it. We [5:00] thought we thought it could be 30%. It's [5:02] hard as hell to measure this. It's not a [5:04] number you can really solidly stand by [5:06] here. And the bottom line is co-pilot is [5:08] not Finn. This is not the AI agent. This [5:11] is not where the value is. Don't settle [5:13] for this damn breadcrumb that is a [5:15] co-pilot. Um Dez, our co-founder, he's [5:18] great at apherisms. He said this down [5:20] here. It starts with augmenting your [5:22] team and co-pilots and everyone's [5:23] telling themselves nothing will really [5:25] change. Aa aka the cope pilot stage. I [5:28] don't know. It doesn't really go well as [5:29] a joke, but anyway, keep going there. [5:31] It's not about co-pilot. What is it [5:32] about? This is the slide we're most [5:34] proud of in R&D and intercom building [5:36] Finn. This is our average resolution [5:38] rate across all of our paying customers. [5:40] That's 6,000 paying customers. And you [5:42] just see over the last two and a half [5:43] years, this going up and up and up and [5:46] up incrementally here. It's the result [5:48] of hundreds of AV tests. This is the [5:49] value we're giving our customers here. [5:51] And this is the actual distribution just [5:53] to try to make this more real. Again, [5:54] across 6,000 customers, there's probably [5:57] no way you can read the numbers at the [5:58] bottom, but you can guess what they are. [6:00] The one in the middle is 60 to 70%. It's [6:02] got like 15% of customers are getting an [6:04] over 80% resolution rate. This is the [6:07] this is the transformation, not your [6:09] co-pilot, the AI agent. This is what's [6:11] on the table. Here's another thing. This [6:14] is from D, our CTO there. There's a [6:15] bunch of stuff we've written up fin.ai [6:18] ideas and he's like the builder buy. So [6:21] this hypothetical, you won't be able to [6:22] gro it fast. I'll try to help you. In [6:24] the middle is Finn with a 50% [6:26] resolution, right? So, even lower than [6:28] it's there at our 99% per res 99 cents [6:30] per resolution. And let's say you decide [6:33] to build it yourself. And let's pretend [6:34] it's free for you to build your own AI [6:37] agent. Obviously, it's not, but let's [6:38] just put that number. It's free. And [6:40] you're able to get a 40% resolution [6:42] rate. You still work out the numbers [6:44] here down at the bottom. You still see [6:45] that 10% resolution gain more than [6:48] compensates for an otherwise free [6:51] product. People underestimate the value [6:53] of resolution rates. That's what I'm [6:54] saying to you. So bottom line, don't [6:56] settle for the breadcrumbs. Don't say we [6:58] tried it. We got some here. We got some [7:00] results and now I have other priorities. [7:01] Like for me as a product person, I'm [7:03] like this has got to be your top [7:05] priority right now. Getting this value [7:07] on the table in your company. Okay, [7:10] that's one. I feel like I'm [7:12] evangelizing. I'll try to do less [7:13] evangelizing up here. Two, you need to [7:15] rethink your metrics. Um I'm going to [7:19] start with here. This is very small, but [7:20] you know, first response time. What the [7:23] hell does that mean? If your first [7:24] response time is immediate anymore, that [7:25] metric just flies out the window. [7:27] Average handling time. Your average [7:30] handling time should go up for humans [7:33] because your AI agent is taking so much [7:35] off their plate. They're dealing with [7:37] the angry customers, the super hard [7:39] customers. That metric is completely [7:41] flipped on its head. Reproductivity is [7:43] super hard to [ __ ] measure anyway, [7:44] right? What are the metrics that measure [7:46] resolution rate, and volume rate, CX [7:48] score? What's CX score? Um so CX score [7:51] is uh a customer experience score sees [7:54] everyone anywhere in CX knows what sees [7:56] right and this is basically this is our [7:58] future AI is not just delivering the [8:00] service AI is measuring the service that [8:02] the service that AI is delivering and [8:04] it's also measuring the service your [8:06] humans are delivering this is our future [8:08] world here so um again there's a long [8:10] blog post I wrote up here explaining on [8:12] this but in short surveys are crazy we [8:15] don't no one responds to surveys and [8:18] we've got the whole transcript there [8:19] it's right for AI to read the whole damn [8:21] transcript of what's going on and [8:23] actually say what's going on uh and [8:25] actually evaluate the quality of that [8:27] conversation instead of relying on a 10% [8:29] response rate. Interestingly, humans [8:31] come out a bit worse and AI agents come [8:33] out a bit better. People like it's [8:35] biased towards itself. Obviously, it's [8:37] not. Anyhow, there's a whole lot more [8:38] here, but we leave there's a whole new [8:40] score here instead of seat a surveyed [8:42] one. It's an AIdriven metric. [8:45] Okay, so I was just telling you how [8:46] proud I am of this metric and I think [8:49] we've been a little too successful of [8:51] getting our customers to anchor on this [8:53] metric because this is actually our [8:55] performance report. We try to give you [8:56] this visual. Here's the flow. Here's [8:58] what's going to f here's what's not [8:59] going to f across all the channels. And [9:01] you see there's two numbers there. I [9:02] talked about two CX score resolution [9:04] rate involvement rate. We actually took [9:06] this off our performance report because [9:07] customer like we don't sorry involvement [9:09] rate. This is how many times did Finn [9:11] have a chance of your inbound volume. [9:13] You got to give Finn a chance. If you [9:14] gave Finn a chance, how many times has [9:16] it successfully resolved that question? [9:18] So, I didn't have to go to your team. [9:19] That's your resolution rate. And what [9:21] happens is CX teams are like not [9:23] thinking about involvement rate. And [9:24] this is really bad because the [9:26] resolution rate like I got a 60% [9:27] resolution rate. That's great. But, uh, [9:29] it's like a 10% involvement rate. And [9:31] you just multiply the two damn numbers [9:33] together and suddenly, oh [ __ ] this is [9:35] not nearly good a picture as you want [9:36] it. This is the research I talked about [9:38] here. They're like, um, uh, teens are [9:40] just not focused on this number. And [9:42] this is a problem because this is our so [9:44] our average of 62% resolution rate our [9:46] average we don't normally share this of [9:48] a 47% involvement rate again our average [9:50] across our paying customers that's a 29% [9:53] automation rate which is both hey a [9:55] third of your all your inbound uh volume [9:58] that's still pretty great start but it's [10:00] a different number than 62%. And that's [10:03] the number that matters most of all. So [10:06] the metrics that matter, resolution, the [10:08] way we kind of think about it, this just [10:09] gives you a bit of the distribution [10:11] reality. The bottom is, we've never [10:13] really shared this before. The bottom is [10:14] like the average Finn automation rate. [10:16] You're still got customers all the way [10:17] over to the very right, but the skew is [10:20] quite a bit different here. So [10:21] resolution rate is really Finn's [10:23] performance. Automation rate is we kind [10:25] of think is your org's performance. I [10:28] kind of buy that. I'm not sure if I buy [10:29] that. Uh but just to get hey even though [10:32] I'm saying resolution rate is not the be [10:33] all and end all in those numbers at the [10:35] start but you still have people [10:36] achieving this transformation. So this [10:39] money group who's getting a 90 98% [10:41] involvement rate you can see the numbers [10:43] here. This is the legitimate proper full [10:45] numbers. Not you need an involvement [10:47] rate and resolution rate. Uh that's what [10:49] matters. Okay breath part three. [10:54] Agents aren't just for the easy stuff [10:56] but it's not easy to get them to do the [10:58] hard stuff. The first thing so many [10:59] people think, yeah, our agents are going [11:00] to do the easy stuff and then the hard [11:02] stuff's for the humans. It's like, okay, [11:04] that's a good first step. Agents are [11:06] going to do the hard stuff. They can do [11:07] it now. But it does take more time and [11:10] it's way easier to get them to do the [11:11] easy stuff. So the way we think about [11:13] this, if all your conversation volume [11:15] that's coming in, we break it into three [11:17] buckets. Informationational queries, [11:18] which is basically the same answer [11:20] regardless of who asked for it, your [11:21] support center is basically given the [11:23] answer to this. Personal queries, which [11:25] is you got to go into a database to get [11:26] that, right? what's my account balance [11:28] or any of the variations there you're [11:29] reading a database and actions where [11:31] you're writing to a database change this [11:33] I need you to change this which is [11:35] obviously the riskiest now the these are [11:38] trying to show ratios here and this [11:39] varies quite a lot per company but you [11:42] also have to remember that the effort [11:44] varies quite a lot per those [11:46] conversation ones so some information [11:47] ones are super quick uh and therefore [11:50] the value can go higher up into the [11:52] harder stuff [11:54] this is a list of like what are the you [11:55] know change change my harder changes [11:57] that this is just a list of a bunch of [11:58] stuff trying to make tangible. What are [12:00] these um actions that you need to do? Um [12:04] okay. And so this is what's you've seen [12:06] a hundred times in these these demos. Oh [12:08] look, the agent can do everything here. [12:10] And it's easy to do this demo. Someone [12:12] was saying this earlier, the dead, [12:13] right? It's easy to do these demos for a [12:15] conversational way that can go in and [12:17] change the order, change this or that. [12:19] You can kind of picture that. What I [12:21] want to touch on a bit and that I think [12:22] is what is less understood is with [12:25] natural language here. So what is this? [12:27] This is a task. This is how you actually [12:29] set this up in Finn behind the scenes. [12:32] You use a natural language trigger. Hey, [12:34] for damaged orders, here's how I want [12:35] you to behave. Use a natural language [12:37] trigger and then you use natural [12:39] language instructions. This is actually [12:41] how you write this out. And the risk [12:43] here is that it seems awfully damn easy. [12:47] And it's deceptive because English is [12:49] now the programming language, [12:52] but you're still doing programming. And [12:54] so what we see is a lot of customers [12:56] shoot themselves in the foot with this. [12:58] It's harder to get right. They take one [13:00] step forward, two steps back. It's just [13:02] hard to do with you need an [13:03] engineering's mindset uh in a lot of [13:05] ways to get this. This will change and [13:07] evolve. But this is why it's harder to [13:09] get this right even though it's natural [13:12] language at its heart doing it is it's [13:14] still a barrier. So I think there's just [13:15] a reality that's worth acknowledging. [13:17] This is like this is from like a two [13:18] months ago an internal uh design like [13:21] why are people struggling with this and [13:22] you can kind of see well one thing [13:24] people haven't written down processes [13:25] it's like you know so much processes in [13:27] people's heads or two they've got to [13:29] learn our system which even though we [13:30] think is relatively easy still a [13:31] learning curve but three that uh [13:34] predictable behavior natural language [13:36] learning basically prompt engineering is [13:38] a skill people on your CS team or CX [13:40] team are going to need to have and then [13:43] finally like internal APIs is a huge [13:45] blocker so right now we've made this [13:47] managed availability because we're like [13:48] we got to help with customers to get [13:50] this right. So the hard stuff is [13:52] absolutely doable today but it is harder [13:54] to get to. Now here what I just wanted [13:57] to show is there is easy stuff that is [13:59] like that deceptively easy sometimes it [14:01] applies. This is a customer I was at [14:02] dinner last week and they were telling [14:04] me on their e-commerce website they're [14:06] having great success with Finn and [14:07] they're like now I'm trying to get it to [14:09] work with leads and they just put this [14:11] and one other like oneliner. This is [14:13] what's called guidance. a simple prompt [14:15] going in and this was he was like I [14:17] couldn't believe it. Finn is now [14:18] actually working successfully doing [14:20] sales uh uh for my leads on the [14:23] e-commerce site. So um sometimes it's [14:26] amazing what you can achieve with just [14:28] oneline prompts in there. So this is [14:29] like guidance which is the simple stuff [14:31] what I showed you before is for complex [14:33] stuff. Okay point4 where am I at for [14:35] time? Five minutes. Jeez, loads of time. [14:38] Okay. Uh CX team roles must change. Uh [14:41] you are now software builders. Uh so [14:44] this is like our system. Everyone's got [14:45] some version of this. It's about [14:46] insights and then it's about train which [14:48] is your content, your guidance, your [14:50] tasks. You got to test this. You're [14:51] basically building software now. You got [14:53] to test this. We're building loads more [14:54] here and then deploy across your [14:56] channels. How are you going to set it [14:57] up? And the point is like what this is [15:00] silly marketing graphic. It's just a [15:02] freaking loop. You got to do this again [15:03] and again and again and again and again. [15:05] It's like building software, right? So [15:08] how does this change this change? This [15:10] is a super busy slide. The stuff on the [15:12] left is like how it used to be. It's all [15:13] about managing humans. The stuff on the [15:15] right, you now need AI specialists. We [15:17] think these are the roles. We're not [15:18] sure what the roles are. We'll keep [15:20] evolving it. This stuff is very wrong. [15:21] And you guys think a Finn is like a huge [15:23] massive teammate. This is the teammate [15:25] you need to build for first because you [15:26] get massive automation here. So, this is [15:28] like the this is the evolution in the [15:31] roles of CX themes that's happening. And [15:34] what are these people doing? So, this is [15:36] um this is a chance to show off some of [15:37] the cool product here, but I won't be [15:39] able to show it. is like, hey, you need [15:40] people diving into the insights. This is [15:41] your all your conversation. This is your [15:43] topics. What is coming and why? [15:45] Everyone's tried to do this. AI unlocks [15:47] way better opportunities to really [15:49] understand your conversation volume. [15:51] What you can see here is we're showing [15:52] this flow. This kind of amazing. This is [15:54] the questions that Finn couldn't answer [15:55] that your team did answer. They seem to [15:58] be informational and we will go in and [16:00] actually write the content we think you [16:02] should change to do that. So, you [16:04] literally click approve. But you need [16:06] people who are focused on this content [16:08] optimization based on what Finn's not [16:10] handling well enough or um or not [16:13] answering there. Um that's just an [16:15] example. This guidance stuff, this is [16:17] like examples of what we use in [16:18] intercom. We have like 80 pieces of [16:20] guidance. This is all fine-tuning that [16:22] we've done over like two years of [16:23] continually getting those numbers better [16:25] and better. You need people focused on [16:27] this not as a project, as a full-time [16:30] ongoing job. [16:32] And then the testing stuff, I won't go [16:34] into that. I just this is this is this [16:36] is not this was shared internally just [16:38] last week of like celebration of our own [16:41] support team and H1 here's the numbers [16:43] they moved you can see a big lift in [16:44] resolution rate involvement rate going [16:46] up some the CX score staying at a good [16:49] place this is what they actually did [16:51] this is what the takes they've been [16:53] building Finn out for our own use case [16:55] here for two years second knowledge [16:57] manager second conversation design all [16:59] around this in getting content ready for [17:02] Finn the first audience I said is get [17:04] your content ready for Finn. This is [17:05] what the new world looks like. Finn is [17:07] like the first audience for there, not [17:08] the agents. Content suggestions, that [17:10] thing I showed you. They're getting like [17:12] a 70% acceptance rate. 28 pieces of [17:15] those oneline fine-tuning. This is what [17:17] the team did and this is what those [17:19] numbers aggraate out. Eight new FIN [17:21] tasks. That's the more complex stuff. [17:22] And the one thing we in R&D can take [17:24] full credit for is like the [17:25] conversational fin product changes where [17:27] we're doing all these AB tests. They get [17:28] that for free. But that's that's what it [17:30] takes. That's what it looks like of what [17:32] a modern uh or future-f facing uh CX [17:34] team is actually doing. Okay. So that's [17:37] the past primarily managing humans and [17:39] tickets. The future primarily building [17:42] and optimizing agents support teams are [17:44] turning into product teams. They're [17:46] building software that is the customer's [17:48] experience. Now the last point here is I [17:51] was talking to some uh some folks [17:52] earlier and they're like hey I got a lot [17:53] of things on on my plate of where to op [17:55] uh where to try to bring AI into where [17:58] to do it first. I'm like, CX is the [18:00] opportunity that's here. I think CX [18:02] should be the pioneer of your company of [18:04] where you're getting massive success. [18:06] Uh, everyone's CX feels like they're in [18:09] the bullseye here. Everyone's in the [18:11] bullseye. We in R&D are bullseye. This [18:13] is an internal thing. We've got a 2x. [18:15] Engineers need to be twice as [18:16] productive. Measure this through PRs. [18:18] What a shitty way to measure engineering [18:20] productivity. We know that. That's how [18:22] we can do it. CX has the best [18:24] measurement for the success of of AI [18:26] agents out there. Now, this is a nice [18:28] quote. Uh I don't actually know who this [18:30] was. Just uh came across it and [18:32] someone's like, "Oh, we're so excited [18:34] about using like anthropic for uh for [18:36] our coding intercom fin." And he's like, [18:38] "We're not just implementing AI tools, [18:40] we're reimagining how humans interact [18:42] with software entirely." The companies [18:44] that figure out this first won't just [18:46] have better support, they'll have [18:47] completely different user experiences. I [18:49] think this is the kind of uh vision that [18:52] people are looking for. And I'll finish [18:53] on this. Consensus talking about AI [18:56] being at the forefront for their [18:57] company. We're leading AI adoption at [19:00] Consensus. People from other teams come [19:02] to us to learn how this works in [19:04] production. We've just shipped this [19:05] literally today to help people map this [19:07] out. Go check it out. A lot of super [19:09] practical details there. CX can be the [19:11] AI pioneer in your company. Thanks, [19:13] folks. I'll be around if anyone wants to [19:15] chat after.